A pilot that sees the problem but cannot afford to act: a rank-check guard at SKY130's fast corner

On 77 SKY130 ff designs a pilot rank check raises the routing shortlist's regret from 2.98% to 4.21%, because falling back to random routing costs too much.

The verified-dse programme asks how much of an open-source chip design search (Yosys, ABC and OpenROAD on SKY130) can be skipped by trusting cheap estimates. Its rule, the clipped ladder (SWC), maps every ABC pass sequence, ranks the mapped netlists with a cheap score, and routes only the best eight. The baseline routes the same number of sequences picked at random, for the same cost.

At the fast-fast corner (ff, -40 C, 1.95 V) SWC still wins on most designs, but it loses badly on some. Three runs looked for a family or delay pattern in those losses, and the last one found none. What the losing designs do share is a cheap score that ranks routed results worse: a median rank correlation (Kendall tau-b) of 0.28, against 0.46 elsewhere.

The guard

So this run tried a per-design check. It routes a few random sequences first, asks how well the cheap score orders them, and if the order is poor, it spends the rest of the budget on random routing instead of the shortlist. The same idea was tried at the typical corner in milestone 111, where it did not help. The fast corner's weaker cheap score gave it a second chance.

The settings grid was 3, 4 or 6 pilot points, and a tau-b cut of -1, 0, 0.2 or 0.4. The unguarded rule was in the grid too. The pool was every finished design from the four fast-corner tables (milestones 144 to 147), 77 designs in all. Settings were chosen on three milestones and scored on the fourth, in turn. Seven predictions were committed before the script first ran. This is a replay of tables that already existed. No new flow ran.

What it found

rulemean regret
unguarded shortlist (SWC)2.98%
guard, settings chosen out of sample4.21%
best guard setting, chosen in sample3.21%
random routing8.95%
  • The guard makes things worse. Out of sample it adds 1.23 points of regret, with a design-bootstrap interval of +0.12 to +3.09. No setting beats the unguarded rule even in sample.
  • The pilot does see the problem. With 6 pilot points and a cut of 0.2, it falls back on 30% of draws where SWC loses and on 18% where it wins.
  • Acting on it is too expensive. On the 14 designs where SWC loses, random routing is 6.28% against SWC's 13.06%, so a correct alarm saves 6.77 points. On the other 59, random is 9.88% against 0.49%, so a false alarm costs 9.39 points. With 14 losers to 59 winners, a guard breaks even only if its false-alarm rate is below 0.17 times its hit rate. This one needed under 5% false alarms and had 18%.
  • Better information would not save it. Even reading the rank correlation over all 120 routed points, which no pilot can do, only gets to 2.82%. That statistic separates SWC's losses from its wins at an AUC of 0.70. The prediction was at least 0.70, and it came in at 0.6998.

A detector that knew in advance which designs SWC loses on would reach 1.75%, so there is a real 1.2 points to win. A rank check on a few random points is not how to win it.

Limits

These tables had been seen before, though the guard's own numbers on them had not been computed. The four folds vary in size, from 6 designs to 31. There is one corner, and routing is global-route estimated. Only one fall-back was tested, plain random routing.

What next

At the typical corner the guard search ended the same way (milestones 126 to 135). The asymmetry points at the fall-back, not the alarm. Random routing is 9.9% on designs where the shortlist is near perfect. A fall-back that keeps part of the shortlist would cut the cost of a false alarm. The next replay tests whether that moves the break-even point on these 77 designs.

Full tables, the grid, the per-design results and the predictions: experiments/pilot_tau_ff/.

All writing · More in Research findings · More on verified-dse · All tools and research